Datasets:
Closure Challenge v2 — CFD case files (heavy)
Companion dataset to anon-closure-challenge-v2/closure-challenge-v2.
This repository hosts the full OpenFOAM cases (mesh, fields, run scripts), VTK volume snapshots, and DNS-native partitioned VTUs for the 14 test cases of the Closure Challenge v2 benchmark, together with the standardized training and validation sets under data/train/ and data/validation/.
The lightweight integral-profile portion needed for reproducing the scoring protocol is in the companion dataset. Most reviewers and submitters do not need this heavy dataset — it is only required for re-running the CFD simulations from scratch or for extended analyses.
NeurIPS 2026 Evaluations & Datasets track submission. Authors anonymised for double-blind review.
Code repository (scoring package, evaluation protocol, per-case attribution): https://anonymous.4open.science/r/closure-challenge-v2
Contents
Approximately 17 GB. 14 test cases, plus 27 training and 5 validation cases. Every case carries a converged solution and the mesh it was computed on.
data/
├── alpha_05_4071_2024/ v1 PHLL 0/, 20000/, constant/, system/, postProcessing/
├── alpha_05_4071_4048/
├── alpha_15_13929_2024/
├── alpha_15_13929_4048/
├── AR_1_Ret_360/ v1 DUCT 0/, 405/, constant/, system/, caseDef, fieldDef
├── AR_3_Ret_360/ 0/, 1540/, ...
├── AR_14_Ret_180/ 0/, 7009/, ...
├── NASA_2DWMH/ v1 hump 0/, 2000/, VTK/, constant/, system/
├── NASA_2DZP/ v2 new 10000/, baseline_komegasst/, highfidelity/
├── NASA_2DN00/ {10,15,17,18}deg/, baseline_komegasst/, highfidelity/
├── NASA_ASJ/ 0/, 0.25/, baseline_komegasst/, highfidelity/
├── NASA_FaithHill/ 0/, 0.orig/, 2000/, VTK/, plots/, ...
├── ERCOFTAC_AhmedBody25/ 0/, 30/, VTK/, plots/, ...
├── ERCOFTAC_WingBodyJunction/ 0/, 5000/, VTK/, plots/, ...
├── high_fidelity_data/ raw upstream reference archives, per case
│
├── train/ 27 TRAINING cases
│ ├── alpha_*/ 21 parametric periodic hills
│ ├── AR_1_Ret_180/ AR_3_Ret_180/ AR_5_Ret_180/ AR_10_Ret_180/
│ ├── CBFS/ curved backward-facing step
│ └── PH_Breuer/ periodic hill, Re=10595
│
└── validation/ 5 VALIDATION cases
├── alpha_05_10071_2024/ alpha_05_10071_4048/
├── alpha_15_7929_2024/ alpha_15_7929_4048/
└── AR_7_Ret_180/
The 14 cases at the top level are the held-out test set. Training and
validation cases live under train/ and validation/ and carry the same
OpenFOAM structure.
Extraction and plotting utilities live at the top of data/:
extract_baseline_from_vtk.py Probe baseline at high-fidelity reference points
extract_wbj_surface_lines.py Wing-body junction surface line extraction
faith_hill_to_csv.py FAITH hill PIV/PSP/FISF re-formatter
commented_dat_to_csv.py Ahmed-format .dat to flat CSV
dat_to_csv.py NASA Tecplot zone-format .dat to flat CSV
Per-case compute_forces.py, extract_surface_Cp.py, and run_hpc_slurm.sh sit inside the relevant case directories.
Standardized training track
The benchmark fixes a train / validation / test split so that methodologies can be compared independently of the training data. All three parts are published here.
data/<case>/ 14 held-out TEST cases
data/train/ 27 training cases
data/validation/ 5 validation cases
| Split | Contents |
|---|---|
| test | 4 periodic hills, 3 ducts, NASA wall-mounted hump, 6 v2 engineering cases |
| training | 21 parametric periodic hills, 4 ducts (AR 1/3/5/10 at Re_tau=180), CBFS, PH_Breuer |
| validation | 4 parametric periodic hills, 1 duct (AR 7 at Re_tau=180) |
Training and validation cases carry the same OpenFOAM structure as the test
cases: constant/polyMesh, system/, initial and converged time directories,
and the high-fidelity reference fields (U_LES, k_LES, tauij_LES).
Using any test case during training or validation is not permitted. The split above is the definition of the standardized track.
Case notes
Reading the periodic-hill case names
The digits in alpha_<slope>_<Lx>_<Ly> are domain dimensions, not Reynolds numbers. alpha_05_4071_2024 has streamwise extent x = 0 .. 4.071 and wall-normal extent y = 0 .. 2.024; alpha_15_13929_4048 spans x = 0 .. 13.929 and y = 0 .. 4.048. The leading alpha value is the hill-slope parameter.
NASA_2DWMH
NASA_2DWMH is the NASA wall-mounted hump (2D), domain x = -0.851 .. 0.672, y = -0.420 .. 0.000, converged at t = 2000. Baseline is simpleFoam + kOmegaSST.
Turbulence model
All 14 baselines are simpleFoam (or rhoSimpleFoam / rhoPimpleFoam for the compressible cases) with RAS kOmegaSST, no ML correction.
License
This dataset aggregates material under two different licenses. See the LICENSE file at the repository root for the authoritative statement.
- CC-BY-4.0: the 12 non-ERCOFTAC test cases (
alpha_*,AR_*,NASA_*) and everything underdata/train/anddata/validation/, derived from public-domain or CC-BY upstream sources (NASA Turbulence Modeling Resource, Vinuesa duct database, Xiao parametric periodic-hill database). - CC-BY-NC-SA-4.0:
ERCOFTAC_AhmedBody25andERCOFTAC_WingBodyJunction, curated extracts of ERCOFTAC kbwiki material (Classic Collection case082; DNS 1-6), inheriting the upstream non-commercial and share-alike conditions. - Code: extraction, plotting, and run scripts (
*.py,*.sh) are MIT. SeeLICENSE-CODEin the code repository.
Hugging Face has no per-directory license field, so this repository declares license: other and records the split in LICENSE. The configs block above groups the two license scopes for readers and machine consumers; it does not itself carry licensing force.
Per-case source attribution is in SOURCES.md of the code repository.
Citation
@inproceedings{closure_challenge_v2_neurips26,
title={The Closure Challenge: A Benchmark Task for Machine Learning in Turbulence Modeling},
author={Anonymous},
booktitle={NeurIPS Datasets and Benchmarks Track (under review)},
year={2026}
}
Croissant metadata
A Croissant 1.0 metadata file is provided as croissant.json in this repository.
- Downloads last month
- 6,242